Using Enhanced Frequency Domain Decomposition as a

Robust Technique to Harmonic Excitation in Operational

Modal Analysis

N-J. Jacobsen
Brel & Kjr Sound & Vibration Measurement A/S
Skodsborgvej 307, DK-2850 Nrum, Denmark
e-mail: njjacobsen@bksv.com

P. Andersen
Structural Vibration Solutions A/S
NOVI Science Park, Niels Jernes Vej 10, DK-9220 Aalborg East, Denmark

R. Brincker
University of Aalborg, Department of Building Technology and Structural Engineering
Sohngaardsholmsvej 57, DK-9000 Aalborg, Denmark

Abstract
The presence of harmonic components in the measured responses is unavoidable in many applications of
Operational Modal Analysis. This is especially true when measuring on mechanical structures containing
rotating or reciprocating parts. This paper describes a new method based on the popular Enhanced
Frequency Domain Decomposition technique for eliminating the influence of these harmonic components
in the modal parameter extraction process.
For various experiments, the quality of the method is assessed and compared to the results obtained using
broadband stochastic excitation forces. Good agreement is found and the method is proven to be an easy-
to-use and robust tool for handling responses with deterministic and stochastic content.

1 Introduction to Operational Modal Analysis

Operational Modal Analysis (OMA) is based on measuring the output of a structure only and using the
ambient and natural operating forces as unmeasured input. It is used instead of classical mobility-based
modal analysis for accurate modal identification under actual operating conditions, and in situations where
it is difficult or impossible to artificially excite the structure.
The algorithms used in OMA assume that the input forces are stochastic in nature. This is often the case
for civil engineering structures like buildings, towers, bridges and offshore structures, which are mainly
loaded by ambient forces like wind, waves, traffic or seismic micro-tremors. The loading forces of many
mechanical structures are, however, often more complex. They are typically a combination of harmonic
components (deterministic signals) originating from the rotating and reciprocating parts and broadband
excitation originating from either self-generated vibrations from, for example, bearings and combustions
or from ambient excitations like air turbulence and road vibrations. However, civil engineering structures
can also have broadband responses superimposed by harmonic components from, for example, ventilation
systems, turbines and generators.
As the input forces to the structure are not measured in OMA, special attention must be paid to identify
and separate harmonic components from true structural modes and eliminate the influence of the harmonic
components in the modal parameter extraction process.
This paper starts by describing the consequences of having harmonic components present in the responses
for different modal parameter identification techniques. Then an overview of various methods for
separating harmonic components from structural modes is given, before explaining in more depth a very
easy-to-use and robust method based on Kurtosis. The well-known and popular Enhanced Frequency
Domain Decomposition (EFDD) modal identification technique is briefly explained, leading to a
description of the new EFDD-based method for eliminating the influence of harmonic components in
OMA.
Finally the quality of the method is assessed from various experiments using a plate structure. The
structure is excited by a combination of a single sinusoidal signal and a broadband stochastic signal and
compared to the results obtained with pure stochastic excitation of the same structure.

2 Consequences of Harmonic Components

The consequences of having harmonic components present in the responses depend on both the nature of
the harmonic components (number, frequency, and level) and the modal parameter extraction method
used. Table 1 below indicates these consequences for the Frequency Domain Decomposition (FDD),
Enhanced FDD (EFDD) and Stochastic Subspace Identification (SSI) techniques. See [1], [2] for further
description of the FDD, EFDD and SSI techniques. The EFDD method is briefly explained in Chapter 4.2
as well.

Technique Consequences
All techniques Harmonic components are potentially mistaken for being structural modes
Harmonic components might potentially bias the estimation of the structural
modes (natural frequency, modal damping, mode shape)
A high dynamic range might be required to extract weak modes
FDD The picked FFT line might be biased by the harmonic component(s)
Harmonics must be away from the structural modes (only the picked FFT line is
used in the FDD technique)
EFDD The identified SDOF function used for modal parameter estimation may be biased
by the harmonic component(s)
Harmonic components must be outside the SDOF function thereby potentially
narrowing the SDOF function and resulting in poorer identification (leakage)
SSI The SSI methods estimate both harmonic components and structural modes. The
(PC, UPC, CVA) modes are estimated correctly even for harmonic components close to - or with
equal frequency as - the modes
Information in the time signal is used both to extract the harmonic components
and the modes, therefore the recording time should generally be longer

Table 1. Consequences of harmonic components for various identification techniques

For the EFDD technique it is important that harmonic components inside the desired SDOF function are
identified and their influence eliminated before processing with the modal parameter extraction process.
It should be noticed that harmonic components cannot, in general, be removed by simple filtering as this
would in most practical cases significantly change the poles of the structural modes and thereby their
natural frequency and modal damping.
3 Identification of Harmonic Components

This chapter gives a brief overview of some useful methods for identifying harmonic components and
structural modes in OMA response data followed by a more detailed description of the Kurtosis method.

3.1 Overview of Methods

In [3] various methods for identifying harmonic components and structural modes were investigated. A
brief overview is presented in Table 2 below.

Technique Description
Short Time Fourier When responses are shown in a contour plot, structural modes are shown as thick
Transform (STFT) vertical lines. Harmonic components are shown as thin vertical lines for stable
conditions (fixed frequencies) and as, for example, saw tooth patterns for run-
up/down conditions (varying frequencies).
Singular Value For a shaped broadband white noise signal exciting the structure, the rank of the
Decomposition matrix containing the singular values will be 1 at frequencies, where only one
(SVD) mode is dominating and higher, if closely-coupled modes or repeated roots are
present. In the case of harmonic components, a high rank will be seen at these
frequencies, as all modes will be excited. The rank will correspond to the number
of responses assuming the dynamic range in the measurement is sufficiently high.
Hence, when the SVD curves are plotted, the peaks will indicate whether they are
due to a harmonic component or a structural mode.
Visual Mode Shapes When a harmonic component is far away from a structural mode, the operating
Comparison deflection shape (ODS) caused by the harmonic component will be a combination
of several excited modes and the loading forces acting on the structure. However,
when a harmonic component is close to an isolated structural mode, the ODS of
the harmonic component will resemble the mode shape and thus can be mistaken
for being a mode shape.
Modal Assurance As a harmonic component will excite all modes, the MAC value between a true
Criterion (MAC) mode shape and an ODS will show high correlation, if the frequency of the ODS is
close to the frequency of the mode shape. On the other hand, the MAC value
between two closely spaced modes will, in general, show low correlation. The
MAC value between two ODS will depend on the modes being excited.
Stabilization Using the SSI techniques, a stabilization diagram showing stable, unstable and
Diagram noise modes is used to select the optimal State Space Dimension. For modes to be
classified as stable, they must fulfil certain mode indicator requirements of which
one is a valid range of damping ratios. By adjusting this range, both harmonic
components and non-physical modes can be filtered out thereby only indicating
the true structural modes as stable modes.
Probability Density The significant difference in the statistical properties of a harmonic component
Functions (PDFs) and a narrowband stochastic response of a structural mode can be used as a
harmonic indicator. Each potential mode is isolated by band-pass filtering, the
PDF is calculated on the result that is subsequently fitted to both the PDF of a pure
harmonic component and the PDF of a pure a structural mode. The prediction
error between the fitted and measured data is calculated in both cases.

Table 2. Overview of methods for identifying harmonic components and structural modes
3.2 Kurtosis

In addition to the above-mentioned methods, kurtosis can also be used as a harmonic indicator. The
kurtosis of a stochastic variable x provides a way of expressing how peaked or how flat the probability
density function of x is. The kurtosis is defined as the fourth central moment of the PDF normalised with
respect to the standard deviation as follows [4]:

(x , ) =
[
E (x )
4
] (1)
4

where is the mean value of x and E is denoting the expectation value.

Often the number 3 is subtracted from equation (1) as this gives a kurtosis of zero, when x follows as
normal distribution [5]:

* (x , ) =
[
E (x )
4
] 3 (2)
4

Using equation (2), a PDF with a positive kurtosis is said to be leptokurtic. If its kurtosis is negative, it is
said to be platykurtic. A PDF with kurtosis equal to zero is called mesokurtic. Leptokurtosis is associated
with PDFs that are simultaneously peaked and have fat tails. Platykurtosis is associated with PDFs
that are simultaneously less peaked and have thinner tails.
For a pure structural mode, the PDF will be normally distributed, and hence the kurtosis * = 0
(mesokurtic). The PDF y is given by:

( x )2
1
y = f (x , ) = e 2 2
(3)
2

A harmonic component can be treated as a stochastic signal if the phase varies randomly within its period.
In the case of a pure harmonic component, the PDF will have two distinct peaks approaching infinity at
a, where a is the amplitude of the harmonic component. The PDF is given by:

y = f (x a ) = ( cos(arcsin ( x a )))
1
(4)
 Figure 1. Normalized PDF of pure structural mode (left) and pure harmonic component (right)

Consequently, the PDF of a harmonic component is associated with leptokurtosis. This fact is used in the
harmonic detection technique described in this paper.

4 Eliminating the Influence of Harmonic Components in EFDD

Before describing the method, the EFDD technique is briefly explained and the consequences of having
harmonic components inside the SDOF function is illustrated.

4.2 Brief Description of the existing EFDD Technique

The Enhanced Frequency Domain Decomposition (EFDD) technique is an extension to the Frequency
Domain Decomposition (FDD) technique. FDD is a basic technique that is extremely easy to use. You
simply pick the modes by locating the picks in SVD plots calculated from the spectral density spectra of
the responses. Animation is performed immediately. As the FDD technique is based on using a single
frequency line from the FFT analysis, the accuracy of the estimated natural frequency depends on the FFT
resolution and no modal damping is calculated. Compared to FDD, the EFDD gives an improved estimate
of both the natural frequencies and the mode shapes and also includes damping.
In EFDD, the SDOF Power Spectral Density function, identified around a peak of resonance, is taken back
to the time domain using the Inverse Discrete Fourier Transform (IDFT). The natural frequency is
obtained by determining the number of zero-crossing as a function of time, and the damping by the
logarithmic decrement of the corresponding SDOF normalized auto correlation function. The SDOF
function is estimated using the shape determined by the previous FDD peak picking - the latter being used
as a reference vector in a correlation analysis based on the Modal Assurance Criterion (MAC). A MAC
value is computed between the reference FDD vector and a singular vector for each particular frequency
line. If the MAC value of this vector is above a user-specified MAC Rejection Level, the corresponding
singular value is included in the description of the SDOF function. The lower this MAC Rejection Level
is, the larger the number of singular values included in the identification of the SDOF function will be.
In the left-hand side of Figure 2, the estimated SDOF function for the plates first bending mode is shown.
A pure SDOF function can be extracted. In the right-hand side of Figure 2, a harmonic component close to
the resonance frequency is present. As seen, the SDOF function is severely affected by the presence of the
harmonic component and hence the estimation of the modal parameters will be incorrect.
 Figure 2. Singular Value SDOF identification without (left) and with (right) harmonic component.
Mode at 354 Hz. Harmonic component at 374 Hz.

Performing an IDFT on the above SDOF functions the normalized correlation functions are calculated. In
the upper part of Figure 3, a typical response is seen of a resonating system that decays exponentially. In
the lower part of Figure 3, the effect of the harmonic component is clearly visible. The harmonic
component can be thought of as a forced vibration with very low damping. The decay is significantly
longer and beating phenomena are observed.

Figure 3. Normalized correlation function without (upper) and with (lower) harmonic component.
Mode at 354 Hz. Harmonic component at 374 Hz. Grey area indicates the part of the function used.

The number of zero-crossings as a function of time calculated from the above correlation functions is
shown in Figure 4. In both cases, the zero-crossing number follows a straight line, but for the case with the
harmonic component, the estimated natural frequency of the mode will be significantly biased by the
presence of the harmonic component and set equal to its frequency.
 Figure 4. Natural frequency identification by zero-crossing counting without (upper) and with (lower)
harmonic component. Mode at 354 Hz. Harmonic component at 374 Hz.

In Figure 5 (next page), the damping ratio is estimated by the logarithmic decrement technique from the
logarithmic envelope of the correlation function. The estimation is performed by applying a linear fit to
the part of the curve being close to a straight line. Again the influence of the harmonic component is
clearly visible.
Compared to the FDD technique, an improved estimate of the mode shape is obtained by using a weighted
sum of the singular vectors i and singular values si whereby random noise is efficiently averaged out.

weight = i si (5)
i

In order not to destroy the mode shape estimate, the singular vector and singular value for the harmonic
component must not be included in the summation.
In Chapter 6, the influence of harmonic components in the modal parameter identification is described
quantitative by various examples.
 Figure 5. Damping ratio estimation from the decay of the correlation function without (upper) and with
(lower) harmonic component. Mode at 354 Hz. Harmonic component at 374 Hz.

4.3 Brief Description of the new EFDD Technique

The steps in the automated method for identifying the harmonic components are roughly as follows:
1. Each measurement channel yi is normalized to unit variance and zero mean
2. For all frequencies fj between DC and the Nyquist frequency a narrow bandpass filtering of yi around fj
is performed
3. The Kurtosis j,i for the filtered signal yi around fj is calculated
4. For each frequency, the mean of the Kurtosis j is calculated across the measurement channels
5. The median m of the Kurtosis of all frequencies is calculated. If the signal is purely Gaussian
distributed this robust measure for the mean will theoretically be 0 (equation (2)).
6. For each frequency fj the deviation of the Kurtosis j from the median m is calculated. If j deviates
significantly from m, then the distribution around fj is different than for the majority of the other
frequencies. In such a case j can be characterised as an outlier that should not be included in the
estimation of the SDOF functions.

In Figure 6 below, the fundamental frequency at 374 Hz as well as the 2nd and 4th harmonics are
automatically identified and shown as vertical green lines in the SVD plot.
 Figure 6. Indication of harmonic components at 374 Hz, 748 Hz and 1496 Hz in the SVD plot.

Knowing the frequencies of the harmonic components, the SDOF function can be estimated by removing
the peaks caused by these harmonic components by using linear interpolation. The global modal
parameters natural frequency and damping can subsequently be calculated. The local parameter
mode shape is calculated as described in equation (5). However, only the non-interpolated singular
values and vectors are used. Figure 7 shows the SDOF function in the SVD plot after removing the
harmonic component.

Figure 7. Removal of harmonic component in SVD plot using linear interpolation.

Mode at 354 Hz. Harmonic component at 374 Hz.

5 Measurement Setup

The measurements were performed using an aluminium plate structure supported by foam rubber as test
object. The plate has lightly damped and well-separated modes. A Brel & Kjr Hand-held Exciter Type
5961 was attached to provide the deterministic signal at a single fixed point. The broadband stochastic
noise input was provided by finger tapping on the plate randomly in time and space to fulfil the criteria for
performing OMA measurements. To avoid mass loading effects across data sets, which is not insignificant
due to the low dynamic mass of the plate, all measurements were done in single data sets by using 16
accelerometers (Brel & Kjr Type 4507-B) equally distributed over the plate. The data acquisition and
analysis was performed using PULSE Type 3560-D front-end and a laptop PC running PULSE
software.
All measurements were done in a 1.6 kHz frequency range using 60 s of time data. A frequency resolution
of 1 Hz was used.

Figure 8. The measurement setup

6 Measurement and Analysis Results

Several tests were conducted to assess the robustness and accuracy of the method as shown in Table 3.

Harmonic Natural Frequency [Hz] Damping Ratio [%]

Comp. [Hz]
EFDD New EFDD EFDD New EFDD
None 354.0 354.0 0.652 0.652
329 329.0 353.8 0.261 0.659
334 334.0 353.9 0.249 0.591
339 339.0 353.9 0.271 0.573
344 344.0 354.6 0.251 0.550
349 349.0 352.4 0.325 1.025
354 354.0 354.3 0.264 1.619
359 359.0 355.0 0.317 1.119
364 364.0 353.7 0.239 0.608
369 369.0 354.3 0.257 0.546
374 374.0 354.0 0.226 0.538
379 379.0 354.3 0.215 0.552

Table 3. Comparison of natural frequency and damping ratio using the existing EFDD
and new EFFD techniques. 1st bending mode at 354 Hz.
Measurements with a single harmonic component located within the SDOF function of the plates 1st
bending mode at 354 Hz were compared to the measurement based on purely stochastic excitation. The
natural frequency and damping ratio were calculated for the existing EFDD technique and for the new
EFDD technique with removed harmonic component.
If harmonic components are present inside the SDOF function, the existing EFDD technique will give
inaccurate estimates of the modal parameters. The natural frequency will be estimated to the frequency of
the harmonic component and the damping ratio is estimated too low.
Using the new EFDD technique the natural frequencies are estimated with a good accuracy and, in
general, the same can be said about the damping ratios. However, when the harmonic component is close
to the natural frequency, larger deviations occur. As the plate is very lightly damped, this is expected due
to the use of linear interpolation and a frequency resolution of 1 Hz. The effect of using a polynomial fit
and a higher frequency resolution will be examined in the near future.
Also the MAC values between the actual mode shape and the estimates obtained after removing the
harmonic component give high correlation (not shown).
Some of the benefits of the new EFDD method are:
Robust method Harmonic components are clearly identified and their effect can be eliminated
even in the case of a harmonic component located exactly at a structural mode. Using high
frequency resolution and/or polynomial fit is required
No prior knowledge required For example about the number of harmonics and their frequencies
Easy-to-use Automated method based on the EFDD technique
Fast Based on computational efficient algorithms

7 Conclusions and Future Work

The presence of dominant harmonic components in the measured responses is unavoidable in many
applications of Operational Modal Analysis. The consequences can be quite drastic, when using the
Enhanced Frequency Domain Decomposition (EFDD) technique, as it requires the harmonic components
to be outside the determined SDOF function. This paper has described a new method based on EFDD
technique, where the harmonic components are first identified using kurtosis and then removed by
performing a linear interpolation across the harmonic components in the SDOF function.
The quality of the method has been assessed from various experiments using a plate structure excited by a
combination of a single sinusoidal signal and a broadband stochastic signal. Compared to the modal
results obtained with pure stochastic excitation of the same structure, the method shows good agreement in
terms of natural frequency, damping ratio and mode shape. Even a harmonic component located exactly at
the peak of a structural mode can be eliminated and a good modal estimate obtained, if a high frequency
resolution is used. Using a polynomial fit, instead of the simple linear interpolation used in this initial
implementation, would improve the results. The method furthermore benefits from not requiring any prior
knowledge of the harmonic components in terms of frequencies or levels, is computational very efficient,
and really simple to use.
Future work will include examination of the methods robustness to multiple and varying frequencies
located within the SDOF function. The effect of the frequency resolution will be investigated and different
polynomial fits will be studied as well. Finally, a new method for eliminating harmonic components
before using Stochastic Subspace Identification (SSI) techniques will be worked on.
References
[1] S. Gade, N. Mller, H. Herlufsen, H. Konstantin-Hansen, Frequency Domain Techniques for
Operational Modal Analysis, 1st IOMAC Conference, 2005

[2] N. Mller, H. Herlufsen, S. Gade, Stochastic Subspace Identification Techniques in Operational

Modal Analysis, 1st IOMAC Conference, 2005

[3] N.J. Jacobsen, Separating Structural Modes and Harmonic Components in Operational Modal
Analysis, IMAC XXIV, 2006

[4] P. Newbold, Statistics for Business and Economics, 2nd edition. Prentice Hall,
ISBN 0-13-845330-6, 1988.

[5] Brel & Kjr, Multichannel Analysis System Type 3550, Vol. 2 Reference and Hardware,
BE1294-11, 1992